Fix tracksFromOPMD Nit_min/Nit_max codepath

[berceanu]

Summary

Two interlocked bugs make the Nit_min/Nit_max arguments of
tracksFromOPMD unusable in the current dev HEAD (7e0d82e):

1. Boolean-mask shape mismatch. The chained-filter form

   iteration_ind = np.arange(iterations.size)
   iteration_ind = iteration_ind[iterations >= Nit_min]   # shrinks
   iteration_ind = iteration_ind[iterations <= Nit_max]   # IndexError

   raises IndexError whenever both bounds are set: the second filter
   applies a full-length boolean mask to the already-shrunken
   iteration_ind. Replaced with a single combined boolean mask,
   which is also clearer.

2. Trajectory arrays never sliced to the window. The slice that
   should follow the iteration filter sits commented out a few lines
   below (# temporal patch to select iterations):

   #for var in var_list:
   #    TC[var] = TC[var][:, iteration_ind]

   ts.iterate walks the full ts.iterations regardless of the
   filter, so each TC[var] returns with shape (N_selected, ts.iterations.size) — but the windowed iterations / t arrays
   built just above suggest the smaller series. Downstream
   it_start indexing inside split_track_by_nans then references
   the wrong column count and produces wrong-length tracks (or
   IndexError's on t[it_start:it_start+z.size] when z_is_xi=True).

   Restored the slice, gated on either bound being set so the
   no-window code path stays untouched, and added a comment explaining
   why it has to live there (after the list-of-lists → ndarray
   conversion).

Context

Hit while integrating SynchRad into a hybrid LWFA→PWFA betatron
pipeline at ELI-NP (FBPIC's BackTransformedParticleDiagnostic →
tracksFromOPMD → SynchRad.calculate_spectrum). Was working
around bug (1) downstream by mutating ts.iterations / ts.t before
the call; this PR makes the upstream API match its docstring so the
workaround can be removed.

Test plan

• Syntax check — python -c "import ast; ast.parse(open('synchrad/converters.py').read())" ⇒ ok

• End-to-end equivalence on a real FBPIC extract. Ran the
  same iteration window twice over a 3001-snapshot
  BackTransformedParticleDiagnostic tree (FBPIC pure-N₂ run, ~1.94M
  candidate particles, sequential sampling of 200):

  • Mode A — stock @ 7e0d82e + the local ts.iterations-mutation
    workaround (Nit_min/Nit_max not passed to tracksFromOPMD).
  • Mode B — PR-patched tracksFromOPMD with Nit_min=2950, Nit_max=3000 passed directly, ts.iterations left untouched.

  Result (200 common track IDs out of 200 written by each):

  check	A	B	Δ
  N_particles	200	200	0
  it_range	[0, 51]	[0, 51]	match
  cdt	6.023597e-06	6.023597e-06	0.00e+00
  shape match (20-track sample)	—	—	20/20
  it_start match (20-track sample)	—	—	20/20
  max |Δ| per field — x, y, z, ux, uy, uz, w	—	—	0.000000e+00 (all)

  PR-patched output is bit-exact with the workaround's output.

• No-window code path (Nit_min=Nit_max=None) — the new gate
  at line 109 (if Nit_min is not None or Nit_max is not None:) only
  fires when at least one bound is set, so the no-window path is
  untouched.

Verification script + raw output available on request.